Calibrating trust in AI-assisted pituitary surgery

Technical Summary: Calibrating Trust in AI-Assisted Pituitary Surgery

Problem Statement
Endoscopic endonasal transsphenoidal surgery (EETS) for pituitary gland tumors presents significant challenges due to the critical nature of neighboring neurovascular anatomy. While Artificial Intelligence Clinical Decision Support Systems (AI-CDSSs) have emerged to assist in real-time anatomical orientation, their safe deployment is hindered by unclear clinician trust dynamics. Trust is not static; it evolves from dispositional to situational and learned forms. In high-stakes surgical environments, inappropriate trust levels—either too low (leading to disorientation) or too high (leading to mission-critical errors)—can compromise patient safety. Previous research suggests that interface design features, such as confidence metrics and explanations, may influence trust calibration, but their specific impact on the dynamic relationship between human trust and AI performance in surgical settings remains unknown.

Methodology
This study was a pre-clinical, IDEAL stage 0 randomized trial conducted online, adhering to DECIDE-AI guidelines. It involved 70 clinicians with pituitary surgery experience (neurosurgeons and medical students), who were randomized 1:1 into two groups:

• Basic AI-CDSS: Provided with an AI-generated overlay outlining the sella turcica on operative video clips.
• Enhanced AI-CDSS: Provided with the same overlay but supplemented with explanations of the model's structure, previous publications, and confidence labels (color-coded: green for >70% reliability, amber for <70%) depicting the reliability of the specific outline.

Experimental Design:

• Stimuli: Participants annotated the sella on six 5-second video clips derived from real EETS operations. The clips were ordered to simulate declining AI performance (DICE coefficients ranging from 0.95 down to 0.68), with the final clip showing intermediate recovery.
• Procedure: Participants first annotated clips without AI assistance, then repeated the task with the AI-CDSS. They could toggle the AI overlay on or off.
• Metrics:
  • Trust: Measured using the Shortened Trust in Automation Scale (STIAS) on a 1-7 Likert scale after each annotation.
  • Performance: Calculated as the DICE overlap coefficient between the participant's annotation and the expert consensus ground truth.
• Analysis: Comparisons between groups utilized Mann-Whitney U tests and permutation analysis. Linear mixed-effects models were used for robustness analysis.

Key Results

• Trust Calibration: When AI performance was high (Clip 1), trust levels were statistically identical between the Basic and Enhanced groups (median 5.00). However, as AI performance degraded, the Enhanced group demonstrated significantly lower trust scores compared to the Basic group. Specifically, at the point of worst AI performance (Clip 5), the Enhanced group's median trust was 3.00 versus 3.67 for the Basic group ($p=.035$). This trend persisted in the final clip ($p=.019$), indicating that the Enhanced interface successfully calibrated trust downward when the AI was unreliable.
• Performance Impact: Overall, the use of AI-CDSS improved participant performance (median DICE increased from 0.83 to 0.85) compared to unassisted annotation. However, there was no significant difference in performance between the Basic and Enhanced groups across the clips.
• Subgroup and Variability Findings:
  • Senior Clinicians: In the best-performing AI scenario (Clip 1), senior clinicians in the Enhanced group showed a trend toward higher performance (median 0.95) compared to the Basic group (median 0.90), though this did not reach statistical significance ($p=.066$).
  • Consistency: When AI was inaccurate (Clip 5), the Enhanced group exhibited significantly less dispersion in performance (IQR 0.07 vs. 0.21, $p=.004$), suggesting the interface promoted more consistent, albeit not necessarily more accurate, annotations during failure modes.

Key Contributions

1. Interface Design and Trust Dynamics: The study provides empirical evidence that modifiable interface features (explanations and confidence labels) can effectively calibrate trust in surgical AI. Specifically, these features help users recognize and react to AI failure more appropriately than a "black box" interface.
2. Safety Mechanisms: The results suggest that enhanced interfaces may serve as safety checks by reducing over-reliance on AI during periods of low accuracy, thereby preventing the "automation bias" often seen in clinical settings.
3. Expert Engagement: The data hints that detailed interface information may encourage senior clinicians to engage more deeply with accurate AI outputs, potentially leveraging their expertise to "cherry-pick" reliable AI suggestions.

Significance and Claims
The authors claim this work is the first to evaluate the impact of interface design on trust within a real-world surgical AI-CDSS context. The study concludes that while interface design does not necessarily alter overall task accuracy in a short-term simulation, it significantly improves trust calibration. By ensuring that trust levels align with actual AI performance, these design features may incrementally improve performance for senior clinicians when AI is accurate and ensure greater consistency when AI is inaccurate. The authors posit that these features are critical for the future safe translation of AI-CDSSs into the operating room, acting as necessary safety checks to mitigate the risks of human-AI team errors in high-stakes neurosurgery.